Meta-learning with normalized projection loss reweighting for webly supervised fine-grained recognition

Clean data is critical to the success of deep learning training. In practice, noisy labels are often included in the data. When the network learns these noisy labels, it will cause the model to degrade. In order to alleviate this problem, the sample reweighting method based on meta-learning has been proposed in recent years. It can reduce the negative impact of noisy labels by adjusting the model’s learning degree to the sample. It also make the model pay more attention to the cleaner data in the dataset. Based on this, we propose Meta-Learning with Normalized Projection Loss Reweighting (MLNP), which is a sample reweighting method based on meta learning. This strategy directs the classification network to identify potentially clean data in the dataset with a higher weight. The weight is based on the Euclidean distance between features and projection similarity of the meta set sample and the training set sample. Furthermore, learning data with less weight is more likely to be noisy. Through experiments, we show the robustness of MLNP and achieves advanced performance on a range of datasets.